Model lab
ForgeLab
Trains small models on one field and proves whether they learned it.
What it does
ForgeLab turns documents, uploaded files or Hugging Face datasets into training data. It strips out duplicates and keeps training and test questions apart, so the results mean something.
It then trains open models on rented GPUs using LoRA and QLoRA, and shows progress as it happens. Once a model is trained, you can chat with it and the original side by side to see what changed. ForgeLab can also distil small models that do one specific job well.
What we use it for
Finding out whether a small model trained on one field can beat a large general one. Our first full experiment, on Oracle E-Business Suite, says it can: a 14B model we trained came out ahead of Claude Opus 4.8 on 289 of 379 questions. The method is on the results page.
This is where Dia's models come from: models that know a customer's systems the way a long-serving analyst does, trained on that customer's own documentation and tested on questions they have never seen.